When Rare Situations Break AI: Researchers Redefine Corner Cases in Machine

Machine-learning systems used in highly automated driving, medical care, and aviation can fail when they encounter rare, new, or dangerous scenarios—so-called “corner cases.” A new open-access study argues that the usual blame on limited or narrow datasets is incomplete. The researchers say model design and architecture also determine when and how prediction errors emerge, and that corner cases should be defined in a way that is measurable, not anecdotal. They note no widely accepted definition exists in the field and that interpretations vary by domain. The paper highlights automotive dataset gaps across regions and conditions, including mislabeling and examples such as a detector confusing a printed image on a taxi banner for a real pedestrian.